All of Nicholas Goldowsky-Dill's Comments + Replies

Yes I'm excited about "probing the answer to followup questions" as a general technique. Our results were promising but there's a lot of iteration one could do to make the technique work better!

 

Separately, did you run any experiments including the probe on the CoT tokens? I would suspect there to be a pretty reliable signal here, and ideally this could be used to help confirm honest CoT. 

On our website you can look at the probe scores on the CoT. I don't have up to date numbers, but I expect it's mildly helpful for improving classification at le... (read more)

One question I have about the Mechanistic Interpretability safety case sketch from Roger Grosse: How do you actually find features of high concern? 

Features of High Concern are things like "intent to carry out a harmful plan". By definition, they very rarely activate on benign deployment outputs. This must be the case as "if the features [of high concern] activate more than a handful of times in deployment (internal or external), we discontinue the model and roll back to a weaker model." [M1.1]

They propose to find these features [S1.1] by pe... (read more)

2RogerGrosse
Great questions. Finding features of high concern is indeed something we highlighted as an open problem, and it's not obvious whether this will be achievable. The avenues you list are plausible ones. It might also turn out that it is very hard to find individual features which are sufficiently indicative of malign behavior, but that one could do so for circuits or specify some other rule involving combinations of features. E.g., the criterion might involve something like "feature activates which indicates the intent to carry out a harmful plan, and features indicating attribution of the thoughts to a fictional character are not active." (As discussed in Limitations, one could apply a variant of this safety case based on circuits rather than features, and the logic would be similar.) We're agnostic about where the model organisms would come from, but I like your suggestions.

Just checking -- you are aware that the reasoning traces shown in the UI are a summarized version of the full reasoning trace (which is not user accessible)?

See "Hiding the Chains of Thought" here.

8RobertM
Yeah, I meant terser compared to typical RLHD'd output from e.g. 4o.  (I was looking at the traces they showed in https://openai.com/index/learning-to-reason-with-llms/).

See also their system card focusing on safety evals: https://openai.com/index/openai-o1-system-card/

Surprising misuse and alignment relevant excerpts:

METR had only ~10 days to evaluate.

Automated R&D+ARA Despite large performance gains on GPQA, and codeforces, automated AI R&D and ARA improvement appear minimal. I wonder how much of this is down to choice of measurement value (what would it show if they could do a probability-of-successful-trajectory logprob-style eval rather than an RL-like eval?). c.f. Fig 3 and 5. Per the system card, METR's eval is ongoing, but I worry about under-estimation here, Devin developers show extremely quick improvem... (read more)

8kolmplex
This system card seems to only cover o1-preview and o1-mini, and excludes their best model o1.

Have you looked at how the dictionaries represent positional information? I worry that the SAEs will learn semi-local codes that intermix positional and semantic information in a way that makes things less interpretable.

To investigate this one can take each feature and could calculate the variance in activations that can explained by the position. If this variance-explained is either ~0% or ~100% for every head I'd be satisfied that positional and semantic information are being well separated into separate features.

In general, I think it makes sense to spe... (read more)

1J Bostock
We might also expect these circuits to take into account relative position rather than absolute position, especially using sinusoidal rather than learned positional encodings. An interesting approach would be to encode the key and query values in a way that deliberately removes positional dependence (for example, run the base model twice with randomly offset positional encodings, train the key/query value to approximate one encoding from the other) then incorporate a relative positional dependence into the learned large QK pair dictionary.

Cool paper! I enjoyed reading it and think it provides some useful information on what adding carefully chosen bias vectors into LLMs can achieve. Some assorted thoughts and observations.

  1. I found skimming through some of the ITI completions in the appendix very helpful. I’d recommend doing so to others.
  2. GPT-Judge is not very reliable. A reasonable fraction of the responses (10-20%, maybe?) seem to be misclassified.
    1. I think it would be informative to see truthfulness according to human judge, although of course this is labor intensive.
  3. A lot of the mileage seem
... (read more)

My summary of the paper:

  1. Setup
    1. Dataset is TruthfulQA (Lin, 2021), which contains various tricky questions, many of them meant to lead the model into saying falsehoods.  These often involve common misconceptions / memes / advertising slogans / religious beliefs / etc.  A "truthful" answer is defined as not saying a falsehoood. An "informative" answer is defined as actually answering the question.  This paper measures the frequency of answers that are both truthful and informative. 
    2. “Truth” on this dataset is judged by a finetuned version of
... (read more)